aiwiki.page
English

Wavelet

A wavelet is a localized oscillatory function used to analyze, represent, and process signals or functions at multiple scales.

39 keywords1 linked from12 not yet writtenWritten by AI
Function (mathem…Fourier Transfor…Harmonic Analysi…Norm (mathematic…Hilbert spaceInner productUncertainty Prin…Orthonormal Basi…Wavelet

A wavelet is a localized, typically oscillatory function whose translated and dilated copies are used to analyze other functions at different positions and scales. Unlike the sinusoidal functions underlying the Fourier transform, wavelets can reveal where a feature occurs as well as the scale on which it varies. A short, sharp pulse, a sudden change in an image, and a slowly varying background therefore produce different patterns of wavelet coefficients.

Wavelets provide both a mathematical theory of multiscale representation and a collection of computational tools. Their foundations connect harmonic analysis, approximation theory, and digital signal processing. Applications include image compression, noise reduction, numerical solution of differential equations, and the analysis of nonstationary measurements. The central distinction is between continuous transforms, which describe variation over a densely sampled range of positions and scales, and discrete transforms, which can provide compact, exactly reconstructible representations.

Historical development

Wavelet theory emerged from several previously separate lines of work. In 1909, Alfred Haar introduced an orthogonal system of piecewise constant functions, published in 1910, that is now recognized as the earliest wavelet basis. The Haar functions are localized and change sign once within their support. They were important in the study of convergence of function expansions, although neither the modern terminology nor the later signal-processing framework existed at the time. Related developments in harmonic analysis included the work of J. E. Littlewood and Raymond Paley during the 1930s, which decomposed functions into frequency bands and related those bands to function norms.

A second line arose from the need to describe signals whose frequency content changes with time. The ordinary Fourier transform represents a signal through global oscillations and does not directly locate individual events. Dennis Gabor’s 1946 work on localized time-frequency analysis used translated and modulated windows. Wavelets eventually offered a different arrangement: instead of keeping the window width fixed while changing frequency, they changed the width together with the oscillation scale. This is especially useful when short-duration high-frequency events coexist with slowly varying low-frequency behavior.

In the late 1970s and early 1980s, the geophysicist Jean Morlet developed scale-dependent analysis for seismic signals. His collaboration with the mathematical physicist Alex Grossmann helped formulate a continuous transform and its reconstruction theory; their influential joint paper appeared in 1984. The French term ondelette, meaning “small wave,” became the source of the English term “wavelet.” Seismic interpretation supplied a concrete motivation: a reflected pulse is localized in time, and its appearance changes with propagation and scale.

During the 1980s, Yves Meyer constructed smooth orthonormal wavelets, while Stéphane Mallat and Meyer developed the multiresolution framework that connected wavelet bases to nested approximation spaces. Mallat’s 1989 work established a systematic connection with fast filter-bank algorithms. Ingrid Daubechies’s 1988 construction of compactly supported orthonormal wavelets showed that exact orthogonality, finite support, and useful regularity could coexist. In 1992, Albert Cohen, Daubechies, and Jean-Christophe Feauveau presented influential constructions of compactly supported biorthogonal wavelets. These developments made wavelets practical for finite digital data as well as useful in analysis. Later advances, including lifting constructions and directional multiscale systems, extended rather than replaced these foundations.

Localization and the continuous wavelet transform

A wavelet is commonly specified by a generating function ψ\psi, called the mother wavelet. Its translated and dilated copies are

ψa,b(t)=1∣a∣ψ ⁣(t−ba),a≠0.\psi_{a,b}(t) = \frac{1}{\sqrt{|a|}} \psi\!\left(\frac{t-b}{a}\right), \qquad a\ne 0.

The parameter bb determines position, while ∣a∣|a| determines scale. Small scales produce narrow wavelets that respond to rapid changes; large scales produce broad wavelets that respond to slower variation. The normalization preserves the norm in the Hilbert space L2(R)L^2(\mathbb R), the space of square-integrable functions. A negative scale also reflects the wavelet. Practical presentations often restrict attention to positive scales, with an appropriate choice of wavelet and reconstruction convention.

For a signal ff, the continuous wavelet transform is the inner product

Wψf(a,b)=⟨f,ψa,b⟩=∫Rf(t)ψa,b(t)‾ dt.W_\psi f(a,b) = \langle f,\psi_{a,b}\rangle = \int_{\mathbb R} f(t)\overline{\psi_{a,b}(t)}\,dt.

A large coefficient indicates a strong match between the signal and the wavelet at that position and scale. Coefficients contain sign or phase information as well as magnitude. Displaying ∣Wψf(a,b)∣2|W_\psi f(a,b)|^2 against position and scale produces a scalogram. Scale is not automatically a physical frequency: its conversion depends on the wavelet’s spectral peak or another chosen center-frequency definition and on the sampling interval.

Exact reconstruction depends on an admissibility condition. With the Fourier convention ψ^(ω)=∫ψ(t)e−iωt dt\widehat\psi(\omega)=\int\psi(t)e^{-i\omega t}\,dt, a standard condition is

0<Cψ=∫R∣ψ^(ω)∣2∣ω∣ dω<∞.0<C_\psi = \int_{\mathbb R} \frac{|\widehat\psi(\omega)|^2}{|\omega|} \,d\omega <\infty.

For sufficiently well-behaved integrable wavelets, this requires ψ^(0)=0\widehat\psi(0)=0, equivalently ∫ψ(t) dt=0\int\psi(t)\,dt=0. The wavelet consequently suppresses a constant background. Using all nonzero real scales, reconstruction takes the form

f(t)=1Cψ∫R∖{0}∫RWψf(a,b)ψa,b(t) db daa2,f(t)= \frac{1}{C_\psi} \int_{\mathbb R\setminus\{0\}} \int_{\mathbb R} W_\psi f(a,b)\psi_{a,b}(t) \,\frac{db\,da}{a^2},

with convergence understood in an appropriate sense, including L2L^2 convergence under standard assumptions.

Localization is limited by the uncertainty principle: a nonzero function cannot be arbitrarily concentrated in both time and frequency. Wavelets distribute this compromise unevenly across frequency. High-frequency components receive fine time resolution and relatively broad frequency bands, while low-frequency components receive coarser time resolution and narrower bands. This approximately constant-relative-bandwidth arrangement distinguishes wavelet analysis from fixed-window methods. It does not imply that wavelets resolve every signal better; their advantage depends on whether the signal’s relevant structures follow this scale-dependent pattern.

Discrete wavelets and multiresolution analysis

The discrete wavelet transform commonly uses dyadic scales and translations:

ψj,k(t)=2j/2ψ(2jt−k),j,k∈Z.\psi_{j,k}(t) = 2^{j/2}\psi(2^jt-k), \qquad j,k\in\mathbb Z.

Here larger jj corresponds to finer scale. Suitable choices of ψ\psi make these functions an orthonormal basis of L2(R)L^2(\mathbb R). A function can then be represented by coefficients dj,k=⟨f,ψj,k⟩d_{j,k}=\langle f,\psi_{j,k}\rangle, with reconstruction

f=∑j,kdj,kψj,k.f=\sum_{j,k}d_{j,k}\psi_{j,k}.

Orthogonality makes the representation nonredundant and preserves squared energy: ∥f∥22=∑j,k∣dj,k∣2\|f\|_2^2=\sum_{j,k}|d_{j,k}|^2. Not every wavelet construction is orthogonal, and not every discrete sampling of a continuous transform gives a basis.

The organizing framework is multiresolution analysis. It consists of nested closed subspaces

⋯⊂V−1⊂V0⊂V1⊂⋯\cdots\subset V_{-1}\subset V_0\subset V_1\subset\cdots

whose union is dense in L2(R)L^2(\mathbb R) and whose intersection is {0}\{0\}. Moving from VjV_j to Vj+1V_{j+1} increases resolution by a factor of two. In a standard orthonormal construction, integer translates of a scaling function ϕ\phi form a basis for V0V_0. The scaling function represents locally averaged or coarse information, whereas the wavelet represents information added when the resolution increases.

The detail space WjW_j is the orthogonal complement of VjV_j within Vj+1V_{j+1}, so that

Vj+1=Vj⊕Wj.V_{j+1}=V_j\oplus W_j.

Iterating gives a coarse approximation plus successive detail layers. For a finite computation ending at resolution JJ, one typically writes

PJf=PJ0f+∑j=J0J−1Qjf,P_Jf = P_{J_0}f+ \sum_{j=J_0}^{J-1}Q_jf,

where PjP_j and QjQ_j are the relevant orthogonal projections. This explains why practical transform outputs contain both a final coarse array and detail arrays. A wavelet’s zero mean does not destroy the overall average: the scaling coefficients retain the coarse information that the detail coefficients omit.

Nestedness leads to a two-scale relation,

ϕ(t)=2∑nh[n]ϕ(2t−n),\phi(t)=\sqrt2\sum_n h[n]\phi(2t-n),

and a corresponding wavelet relation,

ψ(t)=2∑ng[n]ϕ(2t−n).\psi(t)=\sqrt2\sum_n g[n]\phi(2t-n).

The sequences hh and gg become low-pass and high-pass filters. This is the bridge between function-space theory and digital algorithms. It also clarifies a practical distinction: samples of a continuous function are not automatically its scaling coefficients. Software can transform a finite numerical vector directly, but interpreting that vector as a particular continuous-function approximation may require additional sampling or preprocessing assumptions.

Wavelet families and their mathematical properties

The Haar wavelet is the simplest orthonormal example:

ψ(t)={1,0≤t<12,−1,12≤t<1,0,otherwise.\psi(t)= \begin{cases} 1,&0\le t<\tfrac12,\\ -1,&\tfrac12\le t<1,\\ 0,&\text{otherwise}. \end{cases}

Its scaling function is the indicator function of [0,1)[0,1). Haar analysis repeatedly forms normalized averages and differences, making the interpretation unusually direct. The wavelet has compact support and one vanishing moment, but it is discontinuous. It represents piecewise constant behavior efficiently, while smooth curves may require more coefficients than with a smoother family.

The Daubechies wavelets provide compactly supported orthonormal families with increasing numbers of vanishing moments. A wavelet has pp vanishing moments when

∫Rtmψ(t) dt=0,m=0,…,p−1.\int_{\mathbb R}t^m\psi(t)\,dt=0, \qquad m=0,\ldots,p-1.

It therefore gives zero coefficients against polynomials of degree below pp, away from boundary complications. In the standard minimal-support orthonormal construction with pp vanishing moments, the associated finite filters have 2p2p coefficients. Increasing pp generally improves approximation of smooth functions but enlarges support and computational cost. Software naming conventions differ: “D4” commonly denotes a four-coefficient filter, while “db2” denotes the same order through its two vanishing moments.

Other orthogonal families emphasize different compromises. Symlets seek near symmetry within compactly supported orthogonal constructions. Coiflets impose moment conditions on both the wavelet and the scaling function. Meyer wavelets are smooth and band-limited in frequency, but not compactly supported in time. The Shannon wavelet has an idealized sharply bounded frequency support and slowly decaying time-domain tails. These examples illustrate the distinction between finite support, rapid decay, smoothness, and frequency selectivity; none of those properties is interchangeable with the others.

Continuous analysis often uses wavelets chosen for interpretability rather than orthogonal basis construction. The Mexican hat wavelet is proportional to a second derivative of a Gaussian and responds strongly to localized peaks. The Morlet wavelet resembles a complex sinusoid within a Gaussian envelope and supports phase-sensitive analysis of oscillations. The simplest uncorrected expression has a small nonzero mean, so an exact admissible version includes a correction; the correction becomes very small when the carrier frequency is sufficiently high.

Biorthogonal wavelets use distinct analysis and synthesis families that satisfy duality relations rather than requiring a single orthonormal family. This permits compactly supported symmetric filters with linear phase, a useful property in image processing. The price is that coefficient energy no longer obeys the simple orthonormal identity. More broadly, selecting a wavelet requires balancing vanishing moments, regularity, support length, symmetry, spectral behavior, and reconstruction stability. The mathematical properties of the wavelet and its dual both matter when the transform is biorthogonal.

Filter banks, lifting, and finite-data implementation

A fast discrete transform is implemented as a multilevel filter bank. At each level, the input is filtered into low-pass and high-pass branches, and each branch is downsampled by two. In one common convention, approximation coefficients aj[k]a_j[k] produce

aj−1[k]=∑nh[n−2k]‾ aj[n],dj−1[k]=∑ng[n−2k]‾ aj[n].a_{j-1}[k] = \sum_n\overline{h[n-2k]}\,a_j[n], \qquad d_{j-1}[k] = \sum_n\overline{g[n-2k]}\,a_j[n].

Index directions and filter reversals vary among texts and software packages. The essential operations are filtering, decimation, and recursive analysis of the low-pass branch. Reconstruction reverses the process through upsampling and synthesis filtering.

Downsampling creates aliasing in the separate branches. A perfect-reconstruction filter bank is designed so that the aliases cancel when the branches are recombined, restoring the input up to any specified delay or normalization. Thus aliasing within a branch is not by itself a defect. Orthogonal filter banks additionally preserve energy, while biorthogonal banks permit distinct synthesis filters. With a fixed filter length, the work across levels is proportional to

N+N2+N4+⋯ ,N+\frac N2+\frac N4+\cdots,

giving O(N)O(N) complexity in big-O notation for an NN-sample signal. This linear complexity is one of the practical strengths of the dyadic transform.

The lifting scheme, developed prominently by Wim Sweldens in the 1990s, expresses a transform as elementary operations. Samples are split, commonly into even and odd subsequences; one subsequence predicts the other; and an update adjusts the coarse component. Each step has an explicit inverse. Lifting often enables in-place computation, reduces arithmetic, and accommodates nonuniform arrangements. With prescribed rounding rules, it can map integers to integers reversibly, which is important for lossless coding. Such reversible transforms should be distinguished from approximate implementations of a real-valued orthogonal transform.

Finite records introduce boundary questions that the infinite-domain theory does not settle automatically. Periodic extension joins the two ends of the data, zero padding inserts an artificial transition unless the endpoints already vanish, and symmetric extension reflects data near the boundary. These choices change coefficients whose wavelets overlap an endpoint. In continuous-transform displays, the cone of influence marks a region where edge effects become important under a stated convention; it is not a universal guarantee of reliability elsewhere.

The critically sampled transform is also sensitive to translation. Shifting a signal by one sample can substantially rearrange its coefficients because downsampling selects a particular alignment. Undecimated or stationary transforms omit decimation and improve translation behavior at the cost of redundancy. Wavelet-packet transforms instead continue decomposing high-pass as well as low-pass branches, creating finer alternative frequency partitions. Both modify the standard computation, but they address different needs: alignment stability in the first case and adaptable frequency partitioning in the second.

Approximation, sparsity, and statistical estimation

Wavelet coefficients are small in regions where a function is smooth and potentially large near localized irregularities. The mechanism follows from vanishing moments. If ff is sufficiently smooth near a wavelet’s support, its Taylor expansion separates into a low-degree polynomial and a remainder. The polynomial contributes nothing when its degree lies below the number of vanishing moments. Under suitable assumptions, a fine-scale coefficient near a point with regularity exponent ss behaves like

∣dj,k∣≲2−j(s+1/2).|d_{j,k}|\lesssim 2^{-j(s+1/2)}.

The precise bound depends on the function class, wavelet regularity, and location. It expresses why fine-scale coefficients decay rapidly in smooth regions but persist near singularities.

This localization produces sparse representations for many, though not all, signals. In an orthonormal expansion, retaining the MM largest-magnitude coefficients gives a best MM-term approximation in squared L2L^2 error, because omitted squared coefficients add independently. Selection by magnitude is nonlinear: the retained locations depend on the signal. For a function smooth except at a few jumps, most fine-scale wavelets lie entirely inside smooth regions and have small coefficients. Only a limited number at each scale intersect the jumps. This differs from a global Fourier representation, where a discontinuity influences many oscillatory coefficients.

Wavelets also characterize important function spaces. With appropriate regularity and moment conditions, weighted sums of squared coefficients describe norms in Sobolev spaces. More general mixed norms characterize Besov spaces, while related constructions characterize Triebel–Lizorkin spaces. These results turn questions about smoothness and differentiation into questions about sequences. For example, differentiation places greater weight on fine scales because they represent shorter spatial variation. Such characterizations connect wavelet computation to functional analysis rather than merely providing an empirical compression rule.

In statistics, wavelet denoising commonly models observations as a signal plus additive noise. An orthonormal transform preserves the covariance structure of independent Gaussian noise with common variance. Estimated signal coefficients can then be obtained by thresholding. Hard thresholding keeps ww when ∣w∣>λ|w|>\lambda and otherwise sets it to zero. Soft thresholding uses

ηλ(w)=sgn⁡(w)max⁡(∣w∣−λ,0).\eta_\lambda(w) = \operatorname{sgn}(w)\max(|w|-\lambda,0).

David Donoho and Iain Johnstone developed influential risk and adaptivity results for wavelet shrinkage in the 1990s. A widely known threshold is λ=σ2log⁡N\lambda=\sigma\sqrt{2\log N}, derived for a standard Gaussian-noise setting.

Thresholding introduces a trade-off between noise removal and distortion. Soft thresholding shrinks even retained coefficients, while hard thresholding is discontinuous at the cutoff. Correlated noise, redundant transforms, boundaries, and unknown noise levels require modifications to simple theoretical formulas. Connections with regularization are explicit: soft thresholding solves a least-squares problem with an L1L^1 penalty in an orthonormal coefficient system. Sparse coefficients are therefore useful both as a representation and as the basis for a statistical model.

Compression, scientific analysis, and numerical computation

Image compression is a major application because wavelets concentrate much of an image’s information into relatively few large coefficients. A separable two-dimensional transform applies one-dimensional filtering along rows and columns, producing low–low, low–high, high–low, and high–high subbands. These distinguish coarse structure from different kinds of local variation, although horizontal and vertical labels depend on convention. Repeated decomposition of the low–low band produces a hierarchy of resolutions. Compression then combines coefficient quantization with coding of magnitudes, locations, and significance across scales.

JPEG 2000, whose core coding system was standardized in 2000, uses wavelet transforms rather than the block discrete cosine transform associated with baseline JPEG. Its widely used transform choices include the reversible integer 5/3 transform for lossless coding and the irreversible 9/7 transform for lossy coding. The numbers refer to filter lengths in the associated biorthogonal systems. Progressive quality and resolution access depend on the organization of the codestream as well as on the transform itself. Wavelets avoid the fixed small-block structure of baseline JPEG, but heavy quantization can still produce blur or ringing.

Scientific applications exploit localization more directly. In geophysics, wavelets help distinguish seismic events and scale-dependent structure. In turbulence and fluid measurements, they isolate intermittent bursts that global spectra average together. In time-series analysis, continuous wavelet transforms display oscillations whose periods change over time. Cross-wavelet and wavelet-coherence methods compare two records locally in scale and time. Their interpretation requires attention to smoothing, edge effects, noise models, and multiple testing: a visually prominent region of shared power does not establish a causal relationship.

Wavelets are also used to process biological and physical measurements, including electrophysiological signals, astronomical observations, and spectroscopic data. The role is usually computational—separating scales, suppressing specified noise components, or extracting candidate features. Whether those features support a scientific conclusion depends on the measurement process and independently validated analysis. For magnetic resonance imaging, wavelet-domain sparsity can form part of a reconstruction model, particularly when measurements are incomplete. Reconstruction quality then depends on the acquisition operator, noise assumptions, and regularization, not on sparsity alone.

In numerical analysis, wavelet bases provide adaptive representations for partial differential equations and inverse problems. Small coefficients can identify regions needing little resolution, while large fine-scale coefficients identify local complexity. Under suitable assumptions, differential and integral operators have compressible wavelet representations, allowing small matrix entries to be discarded with controlled error. Wavelet coordinates can also support preconditioning and multilevel solvers. Compared with conventional finite element methods, handling irregular geometry and boundary conditions may require more specialized constructions, so the computational advantage is problem-dependent.

Higher-dimensional extensions and structural limitations

The straightforward extension to several dimensions uses tensor products of one-dimensional scaling functions and wavelets. In two dimensions, combinations of ϕ\phi and ψ\psi generate one coarse component and three detail types. This is computationally convenient and well suited to rectangular arrays. However, products aligned with coordinate axes offer only limited directional selectivity. A long diagonal edge can intersect many small wavelet supports, so a representation that is sparse for isolated point singularities need not be equally sparse for curves or surfaces.

Directional multiscale systems address this geometric limitation. Curvelets use elongated elements whose orientation and aspect ratio change with scale; shearlets obtain directional families through anisotropic dilation and shearing. Ridgelets are adapted to line-like structures. These systems are related to wavelets through localization and scale decomposition, but they are not simply alternative one-dimensional mother wavelets. Their approximation properties depend on the geometry assumed for the data. For example, curvelets have strong nonlinear approximation results for idealized images that are smooth away from sufficiently regular curves.

Another extension replaces bases with frames, redundant families satisfying bounds of the form

A∥f∥2≤∑λ∣⟨f,ψλ⟩∣2≤B∥f∥2,0<A≤B<∞.A\|f\|^2 \le \sum_\lambda |\langle f,\psi_\lambda\rangle|^2 \le B\|f\|^2, \qquad 0<A\le B<\infty.

The bounds ensure stable analysis, and reconstruction uses a suitable dual family. Redundancy can improve translation behavior, orientation selectivity, or robustness, but increases storage and computation. The dual-tree complex wavelet transform, developed by Nick Kingsbury, uses parallel real filter trees to approximate complex wavelet behavior. In two dimensions it provides multiple directional subbands and improved shift behavior relative to an ordinary critically sampled separable transform.

Wavelet ideas also extend beyond uniform Euclidean grids. Second-generation constructions use local prediction and update rules on irregular meshes or other nonuniform domains, often through lifting. Spectral graph wavelets use functions of a graph Laplacian to define scale-dependent operators on vertices. Here “scale” concerns the spectral or connectivity structure of a graph, rather than ordinary translation and dilation on the real line. These methods retain a multiscale principle while changing the underlying mathematical machinery.

Several limits remain fundamental across these variants. Finite support conflicts with exact band limitation; strong localization in one domain constrains localization in the other. For real, compactly supported, orthonormal scalar wavelets in the standard dyadic multiresolution setting, exact symmetry is essentially restricted to the Haar case, motivating biorthogonal alternatives. More vanishing moments do not automatically improve every analysis, because longer filters spread the influence of boundaries and nearby events. In higher dimensions, directional coverage and redundancy become additional design choices. A wavelet representation is therefore specified not merely by a family name, but by its sampling, boundary treatment, normalization, reconstruction rules, and intended function or signal class.